Scaling PostgreSQL to power 800 million ChatGPT users
By Bohan Zhang, Member of the Technical Staff
Shell + Skills + Compaction: Tips for long-running agents that do real work
developers.openai.comSaved February 12, 2026
Technical Guide
Summary
The blog shows how to build reliable long‑running OpenAI agents by combining Skills, hosted Shell containers, and server‑side context compaction. It gives concrete patterns for skill schema design, container reuse, network allowlists, and artifact handling.
Highlights
- Skills provide reusable, versioned instruction bundles that can be loaded on demand.
- Hosted Shell runs code in secure, warm containers for continuity across calls.
- Server‑side compaction automatically compresses history to stay under token limits.
- Write precise skill schemas with negative examples and store prompt templates inside skills.
- Reuse the same container, lock down network allowlists, and use /mnt/data for persisted artifacts.
auto-generated
Context
Audience
Developers building production‑grade AI agents and engineers interested in agentic workflow design
DomainGenerative AI Agent Development
Formatblog article with code snippets
Accessfree online
Topics
Related
OpenAI Responses API docsLangChain agentsLLMOps patternsPrompt engineeringContainer security
